Mozilla report: Open Chinese models close gap to US frontier AI

The performance gap between frontier AI models from US tech companies and the best open-weights models from Chinese companies has closed to just 4.4 months, according to a report published by Mozilla. Ars Technica reports that this narrowing gap explains why many companies are shifting to significantly cheaper open models for routine work, while reserving expensive closed models for a narrow band of high-intensity tasks. The report, titled "State of Open Source AI," was published on September 15 and highlights how leading open models are now competitive enough to serve as the default for the majority of organizational workloads.

The details

Mozilla’s report highlights how a leading open model, Moonshot AI’s Kimi K3, achieves a composite AI performance score on the Artificial Analysis Intelligence Index that is just three points behind Anthropic’s Fable 5 closed frontier model. Despite this near-parity in performance, Kimi K3 costs just 30 percent of the latter. Raffi Krikorian, chief technology officer at Mozilla, stated in an email to Ars that a closed model "earns its premium in a few places: expert professional work, high-intensity retrieval, and long context." He emphasized that the decision to pay for closed models should be "workload-specific rather than organization-specific."

The report notes that open-weights models allow anyone to download the main model components and run them on their own computers, though developers typically withhold vital information such as training data and code. In contrast, US tech companies like Anthropic and OpenAI offer closed frontier models that keep everything proprietary. Organizations continue to pay for closed models because they work out of the box and come bundled with "compliance packaging, support, and accountability," features that many organizations lack the staff to replicate with open-weights models.

To measure the narrowing gap, the research nonprofit METR defines an AI model’s time horizon as the length of tasks that can be handled with a reliable 50 percent success rate. Currently, the best closed model can do a job 1.7 times as long as the longest job the best open model can reliably finish. Krikorian explained, "If the open frontier can handle a seven-hour job, the closed frontier can handle a 12-hour one. In four months, the open model handles the 12-hour job, and the closed one handles something around 20." Tasks requiring between eight and 12 hours are typically the ones where closed frontier models still hold an advantage.

Benchmarking company Vals AI evaluated different open and closed models using its own neutral harness to level the playing field. In the Terminal-Bench 2.1 evaluation, the open-weights model GLM 5.2 from Z.ai scored within a point of Anthropic’s Claude Opus 4.7 and 4.8 while costing about five times less per completed task. This suggests that paying for closed frontier models buys about a four-month head start at about five times the per-task cost, but only for tasks taking between eight and 12 hours.

Context

The surging popularity of open-weights models is visible on OpenRouter, an AI gateway and marketplace. The Mozilla report noted that eight of the top 10 models ranked by token volumes in August 2026 provide open weights. However, open models are still lagging behind closed frontier models in revenue. A paper by Frank Nagle and Daniel Yue for the Linux Foundation found that open models earned just 4 percent of overall revenue compared to closed models, which get 96 percent of revenue, based on data from May through September 2025. Krikorian noted, "In the last year, open models have exploded, so we do expect revenue to have shifted."

Krikorian observed that "most of the open models the world runs on are Chinese." He explained that Chinese labs are running the same playbook the Americans ran with Android: "give it away, but own the ecosystem around it." However, he expressed concern about the risk of concentration, noting that the best open models are concentrated in China, whereas the best closed frontier models come from US companies. He argued for US and European labs to start competing in the "same open lane, so that no single country sets the world’s defaults."

What it means for developers

For developers, the practical implication is a shift in default tooling. Routine work that would have previously required expensive closed models can now be handled by open-weights models at a fraction of the cost. Krikorian advised, "Pay when that head start is worth it, something like a deadline that lands before the open frontier catches up would be here. Routine work you’ll still be doing next quarter is not, because you’ll be able to do it for a fifth of the cost soon, and the model won’t be the bottleneck anyway."

Developers should be aware that closed models often come with their own harness, a software layer that helps models access tools and memory. These custom harnesses can boost performance, whereas models may perform less well with third-party harnesses. To get an accurate picture of performance, developers should look for evaluations that use neutral harnesses, such as those from Vals AI, rather than relying solely on vendor-provided benchmarks.

What to watch

• Revenue Shift: Monitor whether the revenue share of open models increases significantly in the coming quarters, as Krikorian expects.
• Geopolitical Concentration: Watch for the emergence of US or European open-model coalitions to counter the current dominance of Chinese labs in the open-weights space.
• Task Horizon Growth: Track the METR time horizon metric to see if the gap between open and closed models continues to narrow or if it stabilizes at the current 4.4-month difference.
• Ecosystem Standards: Look for the development of neutral foundations and public compute programs that fund fully open reference models, as advocated by Krikorian.